AI product photos: what works today and what doesn't

What already works

AI image tools have come a long way, and for a catalog there are three uses that already pay off:

1. Model photos. The product held or worn by a person, without organizing a photo shoot. Useful in fashion, accessories, beauty and any product that is easier to understand in use.
2. Scene photos. The product in context: a table, a workshop, a kitchen. It helps buyers picture the product in their life.
3. Studio versions. A real photo with a clean background and even light, ready for channels that require a white background.

All three have one thing in common: they start from real photos of the product. AI stages what already exists; it does not invent it.

What still fails

Being honest about the limits avoids surprises:

• Small print on the packaging. Labels, ingredients, tiny letters: AI tends to distort them or make up letters.
• Hands. Still among the hardest things, and the first thing to check in a model photo.
• Fine details. Seams, textures, the exact sheen of a finish.
• Inventing what is not in the reference. If the real photos don't show one side of the product, AI imagines it. That is why good references matter.

The practical rule: every generated image is reviewed before it is saved, like any content the buyer will see.

The real photo still rules

A generated photo can sell, but it must not mislead. Buyers must receive what they saw. That is why the main image should be a real photo, or a studio version derived from a real photo without changing the product, and model or scene photos should go in the gallery.

Google recommends high-quality, relevant images with descriptive alt text, placed near the text that explains them. That does not change because an image was generated with AI.

Disclosing what was generated

More and more platforms ask which images were created with AI, and buyers appreciate it too. The most solid way to disclose it is twofold:

• A visible label in your system, so the team knows which images are generated.
• The disclosure inside the file, in its metadata, so it travels with the image wherever it goes.

Google, for its part, says that using AI to create content is not against its rules; what is against them is mass-producing content with the primary purpose of manipulating rankings.

How to give good references

The quality of a generated photo depends mostly on the real photos you give it:

• Sharp, well-lit photos, with the whole product and no cropping.
• Several angles when the product has different sides: front, profile, a close-up of the label.
• Only real photos of that product. A previously generated photo does not work as a reference: it compounds the errors.
• Concrete instructions instead of adjectives: “holding it next to her face” works better than “an elegant photo”.

How Bruxel handles it

In Bruxel, the Studio creates fictional models that are reused across the catalog and generates photos of your products from one to three real photos. The main photo is a studio version derived from a real photo, and it is the only proposal that can become the main image. Each proposal is saved or discarded, and every generated image carries the “Created with AI” label in the app and the disclosure inside the original file. The Studio guide explains it, and the photos for the whole catalog use case shows it on video.

Where to start

Start with the products that today only have a photo on a white background and would be easier to understand in use. Gather their best real photos, generate a few proposals, review them calmly —hands, text, details— and publish only the ones that change nothing about the product.